The problem
A zero upward-step violation rate can be produced by deterministic projection even when the underlying reconstruction is inaccurate. The evaluation must trace that result to its mechanism and compare against simple baselines.
My contribution
- Authorized an AI-assisted audit covering constraint on versus off, matched initialization, held-out cells, and raw versus post-processed outputs.
- Kept the protocol, point predictions, verifier outputs, and claim boundaries reviewable; a personal independent rerun and oral explanation are still pending.
Working approach: product-led, AI-assisted implementation, followed by review and iteration.
Decisions & tradeoffs
01
Trace zero violations to their source
- The choice
- Compare raw, EMA, and EMA-plus-running-min outputs for matched PINN constraint-on/off runs, LSTM, and a no-training noisy-input baseline.
- The tradeoff
- Running-min guarantees no upward steps, but it can preserve an early low value and increase reconstruction error.
- How it is checked
- A separate verifier recomputed RMSE and upward-step violations from saved point predictions; the public links pin the implementation and six CSV snapshots that were audited.
What the work shows
The mechanism audit found that every evaluated method reached zero upward-step violations only after the shared running-min step. Raw outputs still had frequent upward steps, the constraint-on comparison showed no stable gain, and the noisy-input-plus-EMA baseline had the lowest mean reconstruction error. A separate three-seed, six-cell sweep reported PINN 0.9814 versus LSTM 0.2221 Ah mean RMSE across 18 seed-folds, again without support for PINN superiority.
Follow the evidence
What this does not establish
- This protocol reconstructs or denoises observed capacity; it does not forecast future capacity or validate remaining useful life.
- The audit covers six public CSV snapshots, one fixed Gaussian-noise setup, and the inherited implementation. It is not functional-safety certification or deployment evidence.
- The snapshots contain large capacity increases that conflict with a global monotonic assumption; this audit does not establish whether they reflect recovery, test conditions, or extraction error.